Autoencoder with quantification for motor vehicle

Abstract

The invention relates to a system (100) comprising an encoder (110) and a decoder (120), in which a first portion (111) of an autoencoder convolutional neural network is embedded in the encoder and a second portion (122) of the autoencoder convolutional neural network is embedded in the decoder, in which the encoder and the decoder are capable of communicating via a communication channel (130). Potential vectors from a first portion of the neural network are quantized by an encoder for transmission over a communication channel, then dequantized by a decoder, and then processed by a second portion of the autoencoder convolutional neural network.

Description

[0001] The present invention relates to the field of compression of data, in particular image data. The present invention applies in particular but not exclusively to data exchanged in or by a motor vehicle.

[0002] It is known to use an autoencoder system to compress input data. An autoencoder system is a convolutional artificial neural network obtained by performing unsupervised learning on a training data set, comprising an input layer and convolutional layers implemented in an encoder and transposed convolutional layers and an output layer implemented in a decoder, the input layer and the output layer having the same number of nodes and thus the same number of dimensions.

[0003] Thus, the autoencoder system comprises one or more hidden convolutional layers, each hidden layer having a number of dimensions less than the number of dimensions of the input layer and the output layer.

[0004] The hidden layer or central layer with the smallest number of dimensions is called the "code" or "latent vector", and the data represented in this central layer is thus a compressed version of the input data.

[0005] The autoencoder is trained by performing unsupervised learning so as to minimize the mean square error between the input data and the output data originating from the output layer, given the number of dimensions of the central layer or given a compression level.

[0006] Thus, the central layer can be shared between the encoder and the decoder in order to exchange compressed data, thereby making it possible to reduce the data rate requirements and the amount of data exchanged between the encoder and the decoder, while minimizing the losses.

[0007] However, each dimension of the central layer is typically encoded with a high number of bits, in particular a floating-point value of 32 bits. Now, such encoding is not compatible with many data links, in particular with the CAN bus which only accepts 8-bit integer data. For data links other than the CAN bus, it may also be necessary to further compress the data of the central layer.

[0008] To this end, a first aspect of the present invention relates to a method for encoding data, which is implemented in an encoder storing neural layers of an autoencoder convolutional neural network, these neural layers comprising an input layer and at least one hidden convolutional layer, the input layer comprising a first number of dimensions, and said at least one hidden convolutional layer comprising a second number of dimensions, the second number of dimensions being less than the first number of dimensions, the encoding method comprising the following operations:

[0009] - receiving input data of a size equal to the first number of dimensions;

[0010] - Process the input data through a neural layer to obtain a latent vector whose size is equal to the number of the second dimension, and each dimension of the latent vector is encoded in a first N-bit encoding format;

[0011] - Quantize each dimension of the latent vector in a second M-bit encoding format, where M is strictly less than N, to obtain a quantized latent vector;

[0012] - Transmit the quantized latent vector over a communication channel compatible with the second data format.

[0013] Thus, adding quantization enables the quantized latent vector to be transmitted to any communication channel, including channels with a limited data rate or channels that only accept 8-bit data, such as the CAN bus in this case. The quantized vector has a reduced size compared to the input data. In practice, this adaptation causes the compression quality metric of the input data not to degrade or to degrade very slightly.

[0014] A second aspect of the present invention relates to a method for decoding compressed data, which is implemented in a decoder of a neural layer storing an autoencoder convolutional neural network. These neural layers include an output layer and at least one hidden layer. The output layer includes a first number of dimensions, and the at least one hidden convolutional layer includes a second number of dimensions, where the second number of dimensions is less than the first number of dimensions. The decoding method includes the following operations:

[0015] - Receive, over a communication channel, a quantized latent vector whose size is equal to the number of the second dimension. Each dimension of the quantized latent vector is encoded in a second M-bit encoding format, and the communication channel is compatible with the second encoding format;

[0016] - Dequantize the quantized latent vector into a latent vector estimate in the first encoding format, where each dimension of the latent vector estimate is encoded in a first N-bit encoding format, and N is strictly greater than M;

[0017] - Process the latent vector through a neural layer to obtain output data whose size is equal to the number of the first dimension;

[0018] - Transmit the output data.

[0019] According to some embodiments of the first aspect or the second aspect, the communication channel can be a CAN bus, the first N-bit encoding format can be a 32-bit floating-point value, and the second M-bit encoding format can be an 8-bit integer value.

[0020] The CAN bus has the advantage of being a safe and inexpensive link for many devices, especially in motor vehicles.

[0021] According to some embodiments of the first or second aspect, the input data and the output data may represent a motor vehicle lighting photometric map.

[0022] Therefore, an autoencoder convolutional neural network can be used to transmit lighting commands, in particular lighting commands in the form of a lighting photometric map, which is particularly advantageous in motor vehicles.

[0023] According to some embodiments of the first or second aspect, the method may further include a preparatory step of performing unsupervised learning of the autoencoder convolutional neural network on a training data set.

[0024] This machine learning makes it possible to obtain a high compression quality with a very small mean square error between the input data and the output data.

[0025] A third aspect of the present invention relates to a computer program comprising instructions for implementing the method according to the first or second aspect of the present invention when these instructions are executed by a processor.

[0026] A fourth aspect of the present invention relates to an encoder, which includes:

[0027] - A memory that stores the neural layers of the autoencoder convolutional neural network, the neural layers including an input layer and at least one hidden convolutional layer, the input layer including a first number of dimensions, and the at least one hidden convolutional layer including a second number of dimensions, the second number of dimensions being less than the first number of dimensions;

[0028] - A first interface capable of receiving input data having a size equal to the first number of dimensions;

[0029] - A processor configured to: process the input data through the neural layers to obtain a latent vector having a size equal to the second number of dimensions, each dimension of the latent vector being encoded in a first N-bit encoding format; and quantize each dimension of the latent vector in a second M-bit encoding format, where M is strictly less than N, thereby obtaining a quantized latent vector;

[0030] - A second interface capable of transmitting the quantized latent vector on a communication channel compatible with the second data format.

[0031] A fifth aspect of the present invention relates to a decoder, which includes:

[0032] - A memory that stores the neural layers of the autoencoder convolutional neural network, the neural layers including an output layer and at least one hidden convolutional layer, the output layer including a first number of dimensions, and the at least one hidden convolutional layer including a second number of dimensions, the second number of dimensions being less than the first number of dimensions;

[0033] - A first interface that is capable of receiving, over a communication channel, quantized latent vectors equal in size to the number of second dimensions, each dimension of the quantized latent vectors being encoded in a second M-bit coding format, the communication channel being compatible with the second coding format;

[0034] - A processor configured to: dequantize the quantized latent vectors into latent vector estimates in a first coding format, wherein each dimension of the latent vector estimates is encoded in a first N-bit coding format, N being strictly greater than M; and process the latent vectors through a neural layer to obtain output data equal in size to the number of first dimensions;

[0035] - A second interface that is capable of transmitting the output data.

[0036] A sixth aspect of the present invention relates to a system that includes an encoder according to a fourth aspect of the present invention and a decoder according to a fifth aspect of the present invention.

[0037] According to one embodiment, the encoder may be integrated into a central control module of a motor vehicle, and the decoder may be integrated into a lighting device of the motor vehicle.

[0038] Other features and advantages of the present invention will become apparent by reference to the following detailed description and the drawings, in which:

[0039] Figure 1 shows an autoencoder system according to some embodiments of the present invention;

[0040] Figure 2 is a schematic diagram showing steps of a method for processing input data according to some embodiments of the present invention;

[0041] Figure 3 is a schematic diagram showing steps of a method for decoding data according to some embodiments of the present invention;

[0042] Figure 4 shows the structure of an encoder according to some embodiments of the present invention;

[0043] Figure 5 shows the structure of a decoder according to some embodiments of the present invention.

[0044] This description focuses on the features that distinguish the methods, systems, encoders, and decoders from those known in the prior art.

[0045] Figure 1 Shows an autoencoder system 100 according to some embodiments of the present invention.

[0046] ​​​​​System 100 includes an encoder 110 and a decoder 120 connected by a communication channel 130.

[0047] The encoder 110 can be integrated into a device of a motor vehicle, such as a control module responsible for the lighting of the motor vehicle. For example, such a control module can be a PCM (Powertrain Control Module) or an ECU (Electronic Control Unit).

[0048] The decoder 120, for its part, can be integrated into a device of a motor vehicle, such as a lighting device including a lighting module capable of performing a lighting function based on data transmitted by a control module including the encoder 110. Preferably, at least one lighting module of the lighting device is a pixelated module, for example an array having electroluminescent elements (such as LEDs), an array having micromirrors (such as DMDs (Digital Micromirror Devices)), a monolithic source of electroluminescent elements on the same substrate, or any other technology for generating a pixelated lighting beam. The monolithic source involves a plurality of submillimeter-sized electroluminescent semiconductor elements directly epitaxially grown on a common substrate, which is typically made of silicon. Different from a conventional LED array in which each basic light source is a separately produced electronic component mounted on a substrate such as a printed circuit board (PCB), the monolithic source should be regarded as a single electronic component, in the production of which a plurality of ranges of semiconductor electroluminescent junctions are generated in an array form on a common substrate.

[0049] Thus, the communication channel 130 can be a wired link, such as a CAN bus or an Ethernet link. An example of the CAN bus is considered by way of illustration below. It offers the advantages of a secure and inexpensive link. However, the CAN bus requires data to be transmitted in the form of 8-bit encoded integers.

[0050] As a variant, the encoder is integrated into a control module of a motor vehicle, and the decoder 120 is integrated into a server remote from the motor vehicle. In this case, the communication channel includes a wireless communication channel that allows the encoder to access an IP network in which a remote server including the decoder 120 is located. Such a wireless communication channel can be a 3G, 4G, 5G or any next-generation cellular link.

[0051] There is no limitation on the communication channel 130, which can thus be a wired or wireless link. As will be better understood when reading the following description, most communication channels are limited in terms of the data rate of the data they transmit and / or the encoding format.

[0052] According to the present invention, the encoder and the decoder each include a part of an autoencoder convolutional neural network. Thus, the encoder 110 includes a first part 112 of the autoencoder convolutional neural network, while the decoder 120 includes a second part 122 of the autoencoder convolutional neural network.

[0053] As described above, the autoencoder convolutional neural network is obtained by performing unsupervised learning on a training dataset.

[0054] In an example considered here where the encoder is integrated into the PCM or ECU of a motor vehicle and the decoder is integrated into a lighting device, the input data of the autoencoder convolutional neural network is an image representing the light beam to be generated by the lighting device. Such an image is also referred to as a photometric map.

[0055] However, there is no restriction on the input data, which can be any type of image. For example, in an example where the decoder 120 is located in a remote server, the input data can be an image acquired by a vehicle's camera.

[0056] The autoencoder system is capable of compressing such input data. To this end, the autoencoder system is constructed by performing unsupervised learning on a training dataset. The autoencoder system includes an input layer 112 implemented in the encoder 110 and an output layer 124 implemented in the decoder 120, and the input layer and the output layer have the same number of nodes and thus the same number of dimensions.

[0057] The autoencoder system also includes one or more hidden convolutional layers, and the number of dimensions of each hidden convolutional layer is less than the number of dimensions of the input layer 112 and the output layer 124.

[0058] The hidden convolutional layer or the central layer with the smallest number of dimensions is capable of exchanging "codes" or "latent vectors", and thus, the data exchanged between these central layers is a compressed version of the input data.

[0059] Therefore, the central layer can be shared between the encoder 112 and the decoder 123 to exchange compressed data, so that the data rate requirement and the amount of data exchanged between the encoder and the decoder can be reduced while minimizing the loss. Thus, the first part 111 includes an encoding central layer 113, while the second part 122 includes a decoding central layer 123. The encoding central layer 113 and the decoding central layer 123 are capable of exchanging codes or latent vectors including a number of dimensions less than that of the input data. Therefore, the latent vector is a compressed version of the input data.

[0060] The autoencoder is trained by performing unsupervised learning to minimize the mean square error between the input data and the output data derived from the output layer given the number of dimensions of the central layer or at a given compression level.

[0061] For this purpose, the training dataset can be submitted to an autoencoder. The training dataset can include a set of images, such as vehicle lighting luminance maps in the examples considered here. For each image in the training dataset, the autoencoder evaluates the mean squared error between the image submitted to the input layer and the image supplied by the output layer after being processed by the autoencoder, and based on this mean squared error, changes the coefficients of its neurons, the number of neurons, and even the number of hidden convolutional layers, while maintaining constraints to obtain a latent vector with a given number of dimensions. Thus, the aim is to minimize the mean squared error through learning.

[0062] The latent vector is a compressed version of the input data, where the compression ratio CR is calculated according to the following formula:

[0063] CR = (Nbits * Im_Size – NbitsLV * LVdim) / Nbits * Im_Size;

[0064] In the formula, Nbits is the number of bits used to encode each pixel of the input image, Im_Size is the size of the image in terms of the number of pixels, NbitsLV is the number of bits used to encode each dimension of the latent vector, which is fixed and usually equal to 32 bits, and LVdim is the number of dimensions of the latent vector.

[0065] More generally, Nbits * Im_Size represents the size of the input data in terms of the number of bits.

[0066] For a given image size, therefore, the compression ratio CR can be changed by varying the number of dimensions LVdim of the latent vector.

[0067] In particular, the following compression ratios can be obtained:

[0068] - For LVdim = 516, CR = 92%;

[0069] - For LVdim = 1024, CR = 84%;

[0070] - For LVdim = 3072, CR = 52%.

[0071] Therefore, the lower the number of dimensions of the latent vector, the higher the compression ratio CR. The choice of the compression ratio can depend on the quality metric for comparing the output data with the input data. For example, such a metric can include the peak signal-to-noise ratio (PSNR) or the mean squared error (MSE).

[0072] For example, the number of dimensions of the latent vector can be set to ensure a PSNR greater than a given threshold, such as 30.

[0073] Thus, the first part 111 and the second part 122 of the autoencoder convolutional neural network are obtained and can be implemented in the encoder 110 and the decoder 120 respectively. Thus, by arranging a communication channel between the first part 111 and the second part 122, an autoencoder is generated.

[0074] Each dimension of the latent vector is typically a floating-point value encoded in 32 bits (i.e., the first encoding format). Such a value is not compatible with many communication channels, especially with the CAN bus which is only capable of carrying 8-bit integer values.

[0075] To overcome this problem, the present invention proposes to integrate a quantization module 114 into the encoder 110 and a dequantization module 121 into the decoder 120.

[0076] The quantization module 114 is capable of transforming each dimension of the latent vector from a first N-bit encoding format, especially a 32-bit floating-point format, into a second M-bit encoding format, where M is less than N, and in particular, M = 8 for encoding integer values, or as a variant, M = 16. The second encoding format is an encoding format compatible with the communication channel 130.

[0077] In this case, the second encoding format corresponding to 8-bit integer values is compatible with the CAN bus.

[0078] By way of illustration only, an example of quantization is given below in a specific embodiment corresponding to the change from 32-bit floating-point values to 8-bit integer values.

[0079] Suppose the latent vector in the first encoding format according to v = [0.1; 0.2; 0.3] has three dimensions, each dimension being a 32-bit floating-point value between 0 (vmin) and 0.5 (vmax).

[0080] Thus, the quantization module 114 is capable of providing a quantized latent vector having LVdim dimensions, each dimension being encoded in a second encoding format compatible with the communication channel 130.

[0081] The quantized latent vector is obtained as follows:

[0082] Qv = rounded(v / qscale – qzero);

[0083] where rounded() is a function for transforming a floating-point value into an integer value by rounding;

[0084] qscale = vmax – vmin; and

[0085] qzero = 255 – vmax / qscale.

[0086] In the example given above, this thus results in the following quantized latent vector: Qv = [51; 102; 153].

[0087] Therefore, the quantized latent vector Qv can be transmitted via the communication channel 130 to the decoder 120.

[0088] Upon receiving the quantized latent vector Qv, the dequantization module 121 applies the inverse transformation of the quantization module 112 in order to obtain an estimate of the latent vector having LVdim dimensions and each dimension in a first coding format. Thus, the latent vector estimate is supplied at the input of the second part 122 of the autoencoder convolutional neural network in order to determine the output data. Since the convolutional neural network has been trained by machine learning, the output data is close to the input data, although there is a difference between the latent vector originating from the first part 111 and the latent vector estimate originating from the dequantization module 121.

[0089] The addition of the quantization module allows the use of the autoencoder to be compatible with most communication channels, in particular with the CAN bus. Furthermore, while achieving this compatibility, high-quality metrics for input data compression are maintained.

[0090] With the system in Figure 1 it is possible to achieve the following compression levels:

[0091] - For LVdim = 516, CR = 98%;

[0092] - For LVdim = 1024, CR = 96%;

[0093] - For LVdim = 3072, CR = 88%.

[0094] In practice, compression quality metrics such as PSNR and MSE remain at a level as high or almost as high as that of an autoencoder system without quantization.

[0095] Therefore, the system according to the present invention allows the autoencoder convolutional neural network to be compatible with any type of communication channel, in particular with the CAN bus, by increasing the compression ratio while maintaining high compression quality metrics, thus making it possible to minimize the loss of output data.

[0096] Figure 2 Illustrated is a method for processing input data implemented by the encoder 110 described above with reference to Figure 1 described.

[0097] The method includes a preparatory step 200 of storing a first part 111 of an autoencoder convolutional neural network in the memory of the encoder 110. The preparatory step 200 may also include a learning method for obtaining an autoencoder convolutional neural network including the first part 111 and a second part 122.

[0098] The method further includes a step 201 of receiving input data through the input layer 112. As described above, the input data may be data representing an image (such as a photometric map of a motor vehicle lighting device).

[0099] In step 202, the input data is processed by the first part 111, and a latent vector is obtained at the output of the first encoding center layer 113, as described above.

[0100] In step 203, as described above, each dimension of the latent vector is quantized in a second encoding format by the quantization module 114 to obtain a quantized latent vector compatible with transmission on the communication channel 130.

[0101] In step 204, the quantized latent vector is transmitted on the communication channel 130 by the encoder 110.

[0102] Figure 3 Illustrated is a method for decoding data implemented by the decoder 120 described above with reference to Figure 1 description.

[0103] The method includes a preparatory step 300 of storing a second part 122 of an autoencoder convolutional neural network in the memory of the decoder 120. The preparatory step 200 may also include a learning method for obtaining an autoencoder convolutional neural network including the first part 111 and the second part 122.

[0104] In step 301, the dequantization module 121 receives the quantized latent vector. In step 302, the dequantization module 121 determines a latent vector estimate based on the quantized latent vector, and each dimension of the latent vector estimate is in the first encoding format, such as a 32-bit floating-point value. Thus, this step enables a change from the second encoding format to the first encoding format.

[0105] In step 303, the second part 122 of the autoencoder convolutional neural network processes the latent vector estimate to obtain output data, and the output data has the same number of dimensions as the input data from step 201 of the method according to Figure 2 description.

[0106] In step 304, the decoder 120 may transmit the output data. For example, the decoder may transmit the output data to a memory for storage. Advantageously, in an embodiment where the encoder 110 is integrated into the PCM and the decoder 120 is integrated into the signaling device, the output data may be transmitted to the light source control module to implement a photometric map corresponding to the output data.

[0107] The present invention also relates to a method that includes not only Figure 2 the steps of the method, but also Figure 3 the steps of the method, and the receiving step 300 includes receiving the quantized latent vector transmitted in step 204.

[0108] Figure 4 The structure of the encoder 110 according to some embodiments of the present invention is shown.

[0109] The encoder 110 includes a processor 401 that is configured to communicate unidirectionally or bidirectionally with a memory 402 (such as a random access memory (RAM), a read-only memory (ROM), or any other type of memory (flash memory, EEPROM, etc.)) via one or more buses or via a wired connection. As a variant, the memory 402 includes a plurality of memories of the above types. Preferably, the memory 402 is a non-volatile memory.

[0110] The memory 402 permanently or temporarily stores all data generated after implementing steps 201 to 204 of the method for processing data, and may store the first part 111 of the convolutional neural network in step 200.

[0111] The processor 401 is capable of executing instructions stored in the memory 402 to implement steps 202 and 203 of the method shown in reference Figure 2 As an alternative, the processor 401 may be replaced by a microcontroller that is designed and configured to execute steps 202 and 203 of the method according to Figure 3

[0112] Therefore, the quantization module 114 presented above is implemented by the processor 401 or the microcontroller. As another variant, a processor or microcontroller dedicated to the quantization function is provided as a supplement to the processor 401, and this processor thus performs the processing carried out by the first part of the autoencoder convolutional neural network.

[0113] The encoder 110 may include an input interface 403 capable of obtaining input data in the above step 201. There is no limitation on the first input interface 403, and this first input interface may be, for example, a wired interface or alternatively a wireless interface.

[0114] The encoder 110 may also include a second interface, i.e., the output interface 404, capable of transmitting the quantized latent vector in step 204 described above. Thus, the second interface 404 is functionally connected to the communication channel 130.

[0115] Figure 5 The structure of the decoder 120 according to some embodiments of the present invention is shown.

[0116] The decoder 120 includes a processor 501 configured to communicate unidirectionally or bidirectionally with a memory 502 (such as a random access memory (RAM) or a read-only memory (ROM) or any other type of memory (flash memory, EEPROM, etc.)) via one or more buses or via a wired connection. As a variant, the memory 502 includes multiple memories of the above types. Preferably, the memory 502 is a non-volatile memory.

[0117] The memory 502 is capable of permanently or temporarily storing all data generated after performing steps 301 to 304 of the method for decoding data described above, and may store the second part 122 of the convolutional neural network in step 300.

[0118] The processor 501 is capable of executing instructions stored in the memory 502 to implement steps 302 and 303 of the method shown in the reference Figure 3 As an alternative, the processor 501 may be replaced by a microcontroller designed and configured to execute steps 302 and 303 of the method according to Figure 3 Therefore, the dequantization module 121 presented above is implemented by the processor 501 or the microcontroller. As another variant, a processor or microcontroller dedicated to the dequantization function is provided as a supplement to the processor 501, and this processor thus performs the processing carried out by the second part 122 of the autoencoder convolutional neural network.

[0119] The decoder 120 may include an input interface 503 capable of receiving the quantized latent vector in step 301 described above. There is no limitation on the first input interface 503, and this first input interface is functionally connected to the communication channel 130 described above.

[0120] The decoder 120 may also include a second interface, i.e., the output interface 504, capable of transmitting the output data in step 304 described above.

[0121] The present invention is not limited to the embodiments described above by way of example; it extends to other variants.

[0122] The present invention is not limited to the embodiments described above by way of example; it extends to other variants.

Claims

1. A method for encoding data, the method being implemented in an encoder (110) of a neural layer (111) of an autoencoder convolutional neural network, the neural layer including an input layer (112) and at least one hidden convolutional layer (113), the input layer including a first number of dimensions, and the at least one hidden convolutional layer including a second number of dimensions, the second number of dimensions being less than the first number of dimensions, the encoding method including the following operations: - Receiving (201) input data having a size equal to the first number of dimensions; - Processing (202) the input data through the neural layer to obtain a latent vector having a size equal to the second number of dimensions, each dimension of the latent vector being encoded in a first N-bit encoding format; - Quantizing (203) each dimension of the latent vector in a second M-bit encoding format, M being strictly less than N, to obtain a quantized latent vector; - Transmitting (204) the quantized latent vector over a communication channel (130) compatible with the second data format.

2. A method for decoding compressed data, the method being implemented in a decoder (120) of a neural layer (122) of an autoencoder convolutional neural network, the neural layer including an output layer (124) and at least one hidden convolutional layer (123), the output layer including a first number of dimensions, and the at least one hidden convolutional layer including a second number of dimensions, the second number of dimensions being less than the first number of dimensions, the decoding method including the following operations: - Receiving (301) a quantized latent vector having a size equal to the second number of dimensions over a communication channel (130), each dimension of the quantized latent vector being encoded in a second M-bit encoding format, the communication channel being compatible with the second encoding format; - Dequantize (302) the quantized latent vector into a latent vector estimate in a first coding format, wherein, Encoding each dimension of the latent vector estimate in a first N-bit encoding format, N being strictly greater than M; - Processing (303) the latent vector through the neural layer to obtain output data having a size equal to the first number of dimensions; - Transmitting (304) the output data.

3. The method according to claim 1 or 2, wherein The communication channel (130) is a CAN bus, wherein the first N-bit encoding format is a 32-bit floating-point value, and wherein the second M-bit encoding format is an 8-bit integer value.

4. The method according to any one of the preceding claims, wherein, The input data and the output data represent a motor vehicle lighting photometric map.

5. The method according to any one of the preceding claims, further comprising a preparatory step (200; 300) of performing unsupervised learning of the autoencoder convolutional neural network on a training data set.

6. A computer program comprising instructions for implementing the method according to any one of the preceding claims when the instructions are executed by a processor (401; 501).

7. An encoder (110) comprising - A memory (402) that stores neural layers (111) of an autoencoder convolutional neural network, the neural layers including an input layer (112) and at least one hidden convolutional layer (113), the input layer including a first number of dimensions, and the at least one hidden convolutional layer including a second number of dimensions, the second number of dimensions being less than the first number of dimensions; - A first interface (403) that is capable of receiving input data having a size equal to the first number of dimensions; - A processor (401) configured to: process the input data through the neural layers to obtain a latent vector having a size equal to the second number of dimensions, each dimension of the latent vector being encoded in a first N-bit encoding format; and quantize each dimension of the latent vector in a second M-bit encoding format, M being strictly less than N, thereby obtaining a quantized latent vector; - A second interface (404) that is capable of transmitting the quantized latent vector on a communication channel (130) compatible with the second data format.

8. A decoder (120) comprising - A memory (502) that stores neural layers (122) of an autoencoder convolutional neural network, the neural layers including an output layer (123) and at least one hidden convolutional layer (124), the output layer including a first number of dimensions, the at least one hidden convolutional layer including a second number of dimensions, the second number of dimensions being less than the first number of dimensions; - A first interface (503) that is capable of receiving, on a communication channel, a quantized latent vector having a size equal to the second number of dimensions, each dimension of the quantized latent vector being encoded in a second M-bit encoding format, the communication channel being compatible with the second encoding format; - A processor (501), the processor being configured to: dequantize the quantized latent vector into a latent vector estimate in a first coding format, wherein, Encode each dimension of the latent vector estimate in a first N-bit encoding format, N being strictly greater than M; and process the latent vector through the neural layers to obtain output data having a size equal to the first number of dimensions; - A second interface (504) that is capable of transmitting the output data.

9. A system comprising an encoder (110) according to claim 7 and a decoder (120) according to claim 8.

10. The system according to claim 9, wherein, The encoder (110) is integrated into a central control module of a motor vehicle, and wherein the decoder (120) is integrated into a lighting device of the motor vehicle.

Images